Simplicity is the ultimate sophistication. Differentiable Architecture Search (DARTS) has now become one of the mainstream paradigms of neural architecture search. However, it largely suffers from several disturbing factors of optimization process whose results are unstable to reproduce. FairDARTS points out that skip connections natively have an unfair advantage in exclusive competition which primarily leads to dramatic performance collapse. While FairDARTS turns the unfair competition into a collaborative one, we instead impede such unfair advantage by injecting unbiased random noise into skip operations' output. In effect, the optimizer should perceive this difficulty at each training step and refrain from overshooting on skip connections, but in a long run it still converges to the right solution area since no bias is added to the gradient. We name this novel approach as NoisyDARTS. Our experiments on CIFAR-10 and ImageNet attest that it can effectively break the skip connection's unfair advantage and yield better performance. It generates a series of models that achieve state-of-the-art results on both datasets.
We propose to restore old photos that suffer from severe degradation through a deep learning approach. Unlike conventional restoration tasks that can be solved through supervised learning, the degradation in real photos is complex and the domain gap between synthetic images and real old photos makes the network fail to generalize. Therefore, we propose a novel triplet domain translation network by leveraging real photos along with massive synthetic image pairs. Specifically, we train two variational autoencoders (VAEs) to respectively transform old photos and clean photos into two latent spaces. And the translation between these two latent spaces is learned with synthetic paired data. This translation generalizes well to real photos because the domain gap is closed in the compact latent space. Besides, to address multiple degradations mixed in one old photo, we design a global branch with a partial nonlocal block targeting to the structured defects, such as scratches and dust spots, and a local branch targeting to the unstructured defects, such as noises and blurriness. Two branches are fused in the latent space, leading to improved capability to restore old photos from multiple defects. The proposed method outperforms state-of-the-art methods in terms of visual quality for old photos restoration.
We present a general framework for exemplar-based image translation, which synthesizes a photo-realistic image from the input in a distinct domain (e.g., semantic segmentation mask, or edge map, or pose keypoints), given an exemplar image. The output has the style (e.g., color, texture) in consistency with the semantically corresponding objects in the exemplar. We propose to jointly learn the crossdomain correspondence and the image translation, where both tasks facilitate each other and thus can be learned with weak supervision. The images from distinct domains are first aligned to an intermediate domain where dense correspondence is established. Then, the network synthesizes images based on the appearance of semantically corresponding patches in the exemplar. We demonstrate the effectiveness of our approach in several image translation tasks. Our method is superior to state-of-the-art methods in terms of image quality significantly, with the image style faithful to the exemplar with semantic consistency. Moreover, we show the utility of our method for several applications
In deep learning tasks, the learning rate determines the update step size in each iteration, which plays a critical role in gradient-based optimization. However, the determination of the appropriate learning rate in practice typically replies on subjective judgement. In this work, we propose a novel optimization method based on local quadratic approximation (LQA). In each update step, given the gradient direction, we locally approximate the loss function by a standard quadratic function of the learning rate. Then, we propose an approximation step to obtain a nearly optimal learning rate in a computationally efficient way. The proposed LQA method has three important features. First, the learning rate is automatically determined in each update step. Second, it is dynamically adjusted according to the current loss function value and the parameter estimates. Third, with the gradient direction fixed, the proposed method leads to nearly the greatest reduction in terms of the loss function. Extensive experiments have been conducted to prove the strengths of the proposed LQA method.
Single Image Super-Resolution (SISR) aims to improve resolution of small-size low-quality image from a single one. With popularity of consumer electronics in our daily life, this topic has become more and more attractive. In this paper, we argue that the curse of dimensionality is the underlying reason of limiting the performance of state-of-the-art algorithms. To address this issue, we propose Progressive Adversarial Network (PAN) that is capable of coping with this difficulty for domain-specific image super-resolution. The key principle of PAN is that we do not apply any distance-based reconstruction errors as the loss to be optimized, thus free from the restriction of the curse of dimensionality. To maintain faithful reconstruction precision, we resort to U-Net and progressive growing of neural architecture. The low-level features in encoder can be transferred into decoder to enhance textural details with U-Net. Progressive growing enhances image resolution gradually, thereby preserving precision of recovered image. Moreover, to obtain high-fidelity outputs, we leverage the framework of the powerful StyleGAN to perform adversarial learning. Without the curse of dimensionality, our model can super-resolve large-size images with remarkable photo-realistic details and few distortions. Extensive experiments demonstrate the superiority of our algorithm over state-of-the-arts both quantitatively and qualitatively.
Single Image Super-Resolution (SISR) aims to improve resolution of small-size low-quality image from a single one. With popularity of consumer electronics in our daily life, this topic has become more and more attractive. In this paper, we argue that the curse of dimensionality is the underlying reason of limiting the performance of state-of-the-art algorithms. To address this issue, we propose Progressive Adversarial Network (PAN) that is capable of coping with this difficulty for domainspecific image super-resolution. The key principle of PAN is that we do not apply any distance-based reconstruction errors as the loss to be optimized, thus free from the restriction of the curse of dimensionality. To maintain faithful reconstruction precision, we resort to U-Net and progressive growing of neural architecture. The low-level features in encoder can be transferred into decoder to enhance textural details with U-Net. Progressive growing enhances image resolution gradually, thereby preserving precision of recovered image. Moreover, to obtain high-fidelity outputs, we leverage the framework of the powerful StyleGAN to perform adversarial learning. Without the curse of dimensionality, our model can super-resolve large-size images with remarkable photo-realistic details and few distortion. Extensive experiments demonstrate the superiority of our algorithm over existing state-of-the-arts both quantitatively and qualitatively.
As a means of decentralized machine learning, federated learning (FL) has recently drawn considerable attentions. One of the prominent advantages of FL is its capability of preventing clients' data from being directly exposed to external adversaries. Nevertheless, via a viewpoint of information theory, it is still possible for an attacker to steal private information from eavesdropping upon the shared models uploaded by FL clients. In order to address this problem, we develop a novel privacy preserving FL framework based on the concept of differential privacy (DP). To be specific, we first borrow the concept of local DP and introduce a client-level DP (CDP) by adding artificial noises to the shared models before uploading them to servers. Then, we prove that our proposed CDP algorithm can satisfy the DP guarantee with adjustable privacy protection levels by varying the variances of the artificial noises. More importantly, we derive a theoretical convergence upper-bound of the CDP algorithm. Our derived upper-bound reveals that there exists an optimal number of communication rounds to achieve the best convergence performance in terms of loss function values for a given privacy protection level. Furthermore, to obtain this optimal number of communication rounds, which cannot be derived in a closed-form expression, we propose a communication rounds discounting (CRD) method. Compared with the heuristic searching method, our proposed CRD can achieve a much better trade-off between the computational complexity of searching for the optimal number and the convergence performance. Extensive experiments indicate that our CDP algorithm with an optimization on the number of communication rounds using the proposed CRD can effectively improve both the FL training efficiency and FL model quality for a given privacy protection level.
Score matching provides an effective approach to learning flexible unnormalized models, but its scalability is limited by the need to evaluate a second-order derivative. In this paper, we present a scalable approximation to a general family of learning objectives including score matching, by observing a new connection between these objectives and Wasserstein gradient flows. We present applications with promise in learning neural density estimators on manifolds, and training implicit variational and Wasserstein auto-encoders with a manifold-valued prior.
The expressiveness of search space is a key concern in neural architecture search (NAS). Previous approaches are mainly limited to searching for single-path networks. Incorporating multi-path search space with the current one-shot doctrine remains untackled. In this paper, we investigate the supernet behavior under multi-path's setting. We show that a trivial generalization from single-path to multi-path incurs severe feature inconsistency, which deteriorates both supernet training stability and model ranking ability. To remedy this degradation, we employ what we term as shadow batch normalizations (SBN) to catch changing statistics when activating different sets of paths. Extensive experiments on a common NAS benchmark, NAS-bench-101, show that SBN can boost ranking performance at neglectable cost. It breaks the Kendall Tau's record with a clear margin, reaching 0.597. Moreover, we take advantage of feature similarities on activated paths to largely reduce the number of needed SBNs. We call our method MixPath. When proxylessly searching on ImageNet, we obtain several lightweight models that outperform EfficientNet-B0 with fewer FLOPs, parameters and 300x fewer searching resources. Our code will be available https://github.com/xiaomi-automl/MixPath.git .